Theo - t3.ggJune 3, 202620m

Your prompts are tech debt.

TL;DR

Prompts age like code, but fail more quietly: Theo says outdated Agent.md, Claude.md, MCP configs, and system prompt glue can silently degrade model behavior after every model upgrade, unlike normal tech debt that usually breaks loudly.
Bigger prompt setups often make agents dumber: He points to tools like Claude Code reaching about 65,000 tokens of system prompt with features enabled, versus Pi starting under 1,000 tokens, and argues that extra context often wastes capacity and triggers bad tool use.
Prompt tuning is model-specific and expires fast: A prompt that worked for one release can become harmful on the next, which he illustrates with Codex 54 versus 55 and Gemini 3 Pro needing explicit prompt tweaks inside Cursor to stop unnecessary tool calls.
Use stock tools more than bespoke harnesses: Theo agrees with Sean Goedecke that most people should stick close to unconfigured tools like Claude Code, Codex, Cursor, Copilot, or T3 Code, because those teams continuously retest prompts for each model update.
Agent files should contain facts, not hype or ritual phrases: He warns against stuffing agent docs with generic steering like 'think step by step' or AI-generated slop, and recommends keeping them to concrete project details that are actually worth passing every run.
Audit your markdown now: Theo calls out his own T3 Code repo, where an Agent.md was two months stale and still described the project as an early WIP and 'Codex-first,' showing how easy it is for old prompt instructions to survive long after the codebase changes.

The Breakdown

A stale Agent.md can quietly make your AI coding tool worse every time models change, and Theo argues that prompt files should be treated like technical debt you aggressively audit, trim, and delete. His core advice is almost backwards from how many teams work now: customize less, keep context tiny, and let tool vendors do the prompt-tuning grind for each new model.